Image Quality Assessment–driven Reinforcement Learning for Mixed Distorted Image Restoration
Distortion (music)
Representation
DOI:
10.1145/3532625
Publication Date:
2022-04-29T11:37:58Z
AUTHORS (5)
ABSTRACT
Due to the diversity of degradation process that is difficult model, recovery mixed distorted images still a challenging problem. The deep learning model trained under certain declines significantly in other situations. In this article, we explore ways use combination tools deal with distortion. First, illustrate limitations single network dealing multiple distortion types and then introduce hierarchical toolkit distinguished powerful tools. Second, investigate how an efficient representation combined reinforcement (RL) paradigm helps tool noise continuous restoration. proposed method can accurately capture preferences for selecting optimal by RL agent. Finally, fully utilize random unknown combinations, adopt exploration scheme various quality evaluation methods achieve more improvements. Experimental results demonstrate peak signal-to-noise ratio 3.30 dB higher than state-of-the-art RL-based on CSIQ dataset 0.95 DIV2K dataset.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES (40)
CITATIONS (11)
EXTERNAL LINKS
PlumX Metrics
RECOMMENDATIONS
FAIR ASSESSMENT
Coming soon ....
JUPYTER LAB
Coming soon ....